Sonnet 5 Just Outscored Opus on Knowledge Work — at $2 a Million Tokens. Here's the Fine Print.
Anthropic released Claude Sonnet 5 yesterday, June 30. It's the new default model for Free and Pro plans, it's live in the API as claude-sonnet-5, and it ships with an introductory price of $2 per million input tokens and $10 per million output — 40% of what Opus 4.8 costs — through August 31.
Back in February I wrote about Sonnet 4.6 and argued it was the model that made Opus hard to justify for daily work. In April, Opus 4.7 earned its premium in a specific band: long-horizon agent work, security analysis, multi-file discipline. Then Fable 5 opened a new tier above all of it.
Sonnet 5 is the first release in that sequence that moves the boundary from below. On one benchmark it doesn't just approach Opus 4.8 — it edges past it. TechCrunch called it "a cheaper way to run agents," and that's the right frame. But there are three pieces of fine print that change the math, and the introductory pricing is doing some work to hide one of them.
The Numbers
Here's the benchmark picture, compiled across the launch coverage:
| Benchmark | Sonnet 4.6 | Sonnet 5 | Opus 4.8 |
|---|---|---|---|
| SWE-bench Pro (agentic coding) | 58.1% | 63.2% | 69.2% |
| Terminal-Bench 2.1 | 67.0% | 80.4% | — |
| OSWorld-Verified (computer use) | 78.5% | 81.2% | — |
| Humanity's Last Exam (with tools) | 46.8% | 57.4% | 57.9% |
| GDPval-AA v2 (knowledge work, Elo) | — | 1,618 | 1,615 |
Two of those rows matter more than the others.
The Terminal-Bench jump — 67.0% to 80.4% — is the "more agentic" claim made concrete. Terminal-Bench measures whether a model can drive a shell to completion: run commands, read errors, fix, retry. A 13-point single-generation jump on that benchmark is the difference between a model that needs babysitting and one that finishes.
And GDPval, which measures real-world knowledge work, is the first time a Sonnet has scored above its Opus sibling on anything. 1,618 to 1,615 is a rounding error, not a beatdown — but the symbolic line has been crossed. The mid-tier model is no longer strictly dominated.
Early-access partners described the same thing qualitatively: Sonnet 5 finishes complex tasks where 4.6 would stall, and it self-verifies without being told to — one tester described it writing a reproducing test and implementing the fix, unprompted. It's also the first Sonnet-tier model with the xhigh effort level, previously an Opus-only capability. GitHub Copilot had it live on launch day.
Fine Print #1: The Tokenizer Eats Part of Your Discount
Sonnet 5 uses the tokenizer Anthropic introduced with Opus 4.7 — and it produces roughly 1.0 to 1.35x more tokens for the same text than Sonnet 4.6's tokenizer did.
Sticker pricing is unchanged: $3/$15 after the introductory period, exactly what Sonnet 4.6 costs. But you're billed per token, and the same prompt now contains up to 35% more of them. Your real per-task cost on identical workloads can rise by a third while the rate card stays flat.
Three practical consequences:
- Re-baseline with
count_tokensbefore trusting any cost dashboard. Counts measured against Sonnet 4.6 are wrong for Sonnet 5. Don't apply a blanket multiplier — the inflation varies by content type. - Check your
max_tokenslimits. A ceiling tuned for 4.6 output can now truncate equivalent responses mid-thought. Same for compaction triggers keyed to token counts. - The intro pricing masks this until September. At $2/$10, even a 35% token inflation leaves you well below what 4.6 cost. At $3/$15, a worst-case workload costs a third more than it did last week. If your finance team approved "same price as before," they approved the sticker, not the bill.
Fine Print #2: xhigh Effort Can Cost More Than Opus
The xhigh effort level is genuinely new for the Sonnet tier, and for hard coding and agentic tasks it's where the benchmark numbers live. But effort is a token multiplier — higher effort means more thinking, more tool calls, more self-verification passes, all billed as output.
The launch analysis found that at xhigh effort, Sonnet 5's per-task cost can exceed Opus 4.8's at similar quality. Read that again: the cheap model, pushed to its ceiling, can cost more than the expensive model coasting.
The rough cross-model mapping from Anthropic's own migration guidance: Sonnet 5 at medium is comparable to Sonnet 4.6 at high, and Sonnet 5 at high matches Sonnet 4.6 at max. So the cost-efficiency sweet spot is low and medium effort — routine work where Sonnet 5 delivers previous-generation-Opus quality at Sonnet prices. If a task genuinely needs xhigh, run the math against Opus 4.8 at high first. You might be paying Sonnet prices for Opus bills.
Fine Print #3: It's a Breaking Change, Not a Drop-In
If you're calling the API, Sonnet 5 aligns the Sonnet line with the Opus 4.7/4.8 request surface — which means code written for 4.6 can silently change behavior or outright 400:
- Adaptive thinking is on by default. On Sonnet 4.6, omitting the
thinkingparameter ran without thinking. On Sonnet 5, the same request runs with adaptive thinking — spending thinking tokens you weren't spending before, inside amax_tokensbudget you sized without them. Setthinking: {type: "disabled"}explicitly if you want the old behavior. budget_tokensnow returns a 400. The transitional escape hatch that still worked on 4.6 is gone. Adaptive thinking pluseffortis the only path.- Non-default sampling parameters are rejected.
temperature,top_p,top_k— remove them and steer with prompting. - Thinking display defaults to omitted. If you stream reasoning to users, the default now looks like a long pause before output. Set
display: "summarized"explicitly.
None of this is surprising if you migrated anything to Opus 4.7 or 4.8 — it's the same surface. But "same model family, same price" makes it tempting to just swap the model string, and that's exactly the workload that gets bitten.
The Security Read
Sonnet 5 ships with the same default-on cybersecurity safeguards as Opus 4.7, 4.8, and Fable 5 — and Anthropic's system card is unusually blunt about the capability line: Sonnet 5 scored 0% on developing working exploits in their Firefox vulnerability tests, by design substantially weaker than Opus 4.8 on offensive cyber tasks.
For defenders, the safety profile actually improved in useful ways: lower misaligned-behavior rates than 4.6, better refusal of malicious requests, stronger prompt-injection resistance — which matters a lot for a model whose headline improvement is autonomous computer use at 81.2% on OSWorld. An agent that clicks through the open web needs exactly that hardening.
The advice I gave for Opus 4.7's safeguards carries over unchanged: if you do authorized security work, benign-adjacent tasks will occasionally trip false positives. Build the fallback path into your integration before you hit it in production.
The Updated Decision Table
Here's where the February argument lands now:
- Sonnet 5 at low/medium effort is the new default for everything Sonnet 4.6 was the default for — plus a real chunk of what used to need Opus. Agentic coding loops, computer use, document work, multi-step automation. This is most of your volume.
- Opus 4.8 keeps a narrower band: accuracy-critical work where a subtle miss is expensive, tasks that would force Sonnet 5 to
xhighanyway (where the cost advantage inverts), and offensive-adjacent security analysis where Sonnet 5 is deliberately capability-limited. - Fable 5 remains the answer for the frontier: the multi-day autonomous runs and team-scale migrations I covered at launch.
The one-line verdict: Sonnet 5 makes Opus 4.8 genuinely optional for most workloads — the first Sonnet where that's an accurate description rather than a marketing line. But "optional" was earned with a tokenizer that quietly claws back up to a third of the savings, and an effort ceiling that costs Opus money at Sonnet quality. Take the $2/$10 window through August to measure your actual per-task costs. That's what the discount is for.